acceptodds
Withdrawn from ICLR 2027

FedCrossID: Recovering Fragmented Cross-Camera Supervision in Federated Person Re-Identification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Person re-identification (ReID) relies on recognizing the same person across cameras. In federated-by-camera training, however, each camera uses an independent identity namespace, leaving cross-camera correspondences unobserved. Existing federated ReID methods improve aggregation, domain generalization, or local feature diversity, but generally leave unknown correspondences between partially overlapping client identity sets unused. Two difficulties arise: client identity sets overlap only partially, and uncertain inferred relations can exert increasing influence relative to the main-task gradient. We introduce FedCrossID, a discover-then-calibrate framework. Protected Partial Identity Association (PPIA) discovers soft correspondences with an unmatched option, while Reliability-Calibrated Relation Learning (RCRL) weights detached relation targets by confidence and attenuates their influence over communication rounds. Under a protocol that withholds global identity mappings from training, FedCrossID improves a local FedPav-style baseline by 16.81 and 4.17 mAP points on Market1501 and MSMT17, reaching 49.22 ± 0.52 and 17.33 ± 0.14, respectively. Across three paired seeds, PPIA provides most of the improvement. Loss-formulation controls attribute most of the additional MSMT17 gain to the revised relation-loss formulation. Gradient diagnostics motivate phase attenuation to control the changing relative influence of relation supervision. In the tested settings, selected relations outperform degree-preserving randomized relations, including at zero identity overlap; this does not establish that the gains come from recovering true identities. Tuned constants achieve similar retrieval, and partial OT does not consistently outperform forced matching. Summary perturbation provides no formal privacy guarantee.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Person Re-Identification, Cross-Camera Identity Association, Confidence-Weighted Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bVjSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FedCrossID studies federated person re-identification when each camera has an independent identity namespace and cross-camera identity mappings are unavailable to the normal training procedure. Protected Partial Identity Association (PPIA) uses optimal transport to match perturbed local-identity summaries, adjusts matching scores according to uncertainty, and allows identities to remain unmatched across clients. Reliability-Calibrated Relation Learning (RCRL) uses remote prototypes associated with accepted relations as supervision targets without backpropagating through those targets. Confidence weights and a round-dependent coefficient control the relation loss. Experiments on Market1501 and MSMT17 supports the utility of selected cross-client relations.

Critical strengths
  1. Independent camera-local identity labels make cross-client relations difficult to use directly in local supervision. FedCrossID introduces candidate associations into training and treats relation discovery and utilization separately, providing the shared encoder with cross-camera supervision beyond local identity classification. This offers a plausible approach to label fragmentation in federated ReID.

  2. PPIA performs partial matching use identity summaries and their uncertainty, with an unmatched option and mutual filtering to accommodate partially overlapping client identity sets. RCRL converts accepted associations into supervision based on remote prototypes and regulates its strength through confidence weights and a round-dependent coefficient. The two modules address which identities to associate and how to use these uncertain associations, aligning the method's design with the stated challenges.

Critical weaknesses
  1. Partial optimal transport does not consistently outperform forced matching. Compared with validation-selected fixed coefficients, coefficient decay improves mAP by only 0.23 ± 0.54 and 0.26 ± 0.24 percentage points on the two datasets. Furthermore, the ablation study on the loss function involves simultaneous changes to multiple factors. Consequently, the existing evidence provides stronger support for the effectiveness of the overall approach than for the independent contributions of its individual components.

  2. The authors employed "global identity mappings" to determine the association temperature. The performance of the method when parameter selection is conducted without these mappings has not been evaluated. Furthermore, the experimental conditions used for the published results cited in Table 1 differ from those employed in the authors' own implementation, making it difficult to conclude definitively that this method offers an advantage over stronger baselines.

Questions for authors
  1. Is it possible to obtain retrieval results for temperatures near the current value (or for specific selected temperatures without relying on global identity mappings)? Such results would help clarify the system's sensitivity to temperature, as well as the extent to which parameter selection depends on the identity mappings established during the development phase?

  2. In scenarios involving identity overlap, are there statistics available regarding the precision and recall of the adopted associations, and how do these metrics evolve over the course of training epochs? Furthermore, how do these metrics shift in response to changes in retrieval performance?

  3. Switching from the legacy loss function to RCRL alters the aggregation of positive targets, the sample weighting scheme, and the normalization process simultaneously. Is there a control setup that allows for replacing confidence-based weights with uniform weights while keeping all other configurations constant? If not, what conclusions regarding confidence-based weighting can be drawn from the existing results, and which contributions remain inseparable?

Reviewer_7zB7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

FedCrossID addresses federated-by-camera person ReID with independent local identity labels. It matches identity prototypes across clients using perturbed summaries and dustbin-augmented optimal transport, then trains with confidence-weighted cross-client targets and a decaying relation-loss coefficient. The method improves retrieval over the locally implemented baseline on Market1501 and MSMT17.

Critical strengths
  1. The controlled-overlap experiments and degree-preserving randomization examine whether gains depend on identity overlap and relation selection, rather than merely adding more training edges.

  2. The controlled-overlap experiments and degree-preserving randomization examine whether gains depend on identity overlap and relation selection, rather than merely adding more training edges.

Critical weaknesses
  1. The problem motivation substantially overlaps with established ReID settings. Independent camera-local labels and missing cross-camera correspondences are central to intra-camera supervised ReID. MTML explicitly discovers correspondences across camera-specific label spaces[1]. Camera-incremental methods also address unknown identity overlap[2]. Although their access patterns differ from federated training, the underlying motivation and several methodological ideas are closely related.

  2. The experiments do not establish identity recovery as the source of improvement. The retrieval gains persist even at zero identity overlap, where none of the accepted links can represent the same person. This suggests that the method may benefit from linking visually similar identities rather than recovering true correspondences. The authors acknowledge this, but the paper’s main explanation remains unresolved. Reporting association precision and recall, together with an oracle-link comparison, would help clarify what drives the gains.

  3. The ablations provide limited support for the two proposed components: partial OT does not consistently beat forced matching, while the decaying coefficient performs similarly to a tuned constant. Moreover, the weighting comparison changes several aspects of the loss at once. It is therefore unclear how much the full design adds beyond a simple prototype matcher trained with the revised relation loss.

  4. The improvement over the local baseline is encouraging, but Table 1 does not establish how the method compares with existing approaches under the same conditions. The published results come from a different pipeline, and the reproduced FedPav baseline differs substantially from its reported performance. Without matched comparisons against stronger federated and association-based methods, the practical advantage remains difficult to assess.

[1]Towards Precise Intra-Camera Supervised Person Re-Identification. WACV, 2021. [2]Handling Label Uncertainty for Camera Incremental Person Re-Identification. ACM Multimedia, 2023.

Questions for authors

What measurable advantage remains over a simple prototype matcher with the final relation loss and a tuned constant on the full datasets?

Reviewer_5ebHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates cross-camera supervision in federated Person re-identification (ReID) and proposes FedCrossID. Authors consider a federated-by-camera training setting in which each client uses an independent identity space and cross-camera indentity correspondences are unavailable during normal training. The proposed methods combines Protected Partial Identity Association (PPIA), which infers candidate correspondences using perturbed prototypes and partial optimal transport with an unmatched option, and Reliability-calibrated Relation Learning (RCRL), which incorporates these relations through detached targets, confidence weighting, and a round-dependant coefficient. Experiments on Market1501 and MSMT17 demonstrate improvements in FedPav-style baseline.

Critical strengths

S1: Camera-based local discripencies are well motivated

  • The motivation is important and clearly presented. Independent camera's local identity labels create a large supervision gap, and using cross-client relations is a reasonable approach.

S2: Intuitively convicing suggestion

  • The proposed approach FedCrossID is intuitive.
Critical weaknesses

W1: Main presentation needs big improvement

  • The presentation of Figure 1 needs substantial improvement. Figure 1 is intended to explain the complete framework, but its mathematical symbols are inconsistent with the main text. For example, the relation loss is written as "Lrel" rather than Lrel\mathcal{L}_{rel}, and other labels, including the matrix and uncertainty terms, do not preserve the corresponding mathematical notation. Moreover, the placement and alignment of several arrows in the PPIA part also make the information flow difficult to follow.

W2: Limited rationale for combining two key parts

  • Authors acknowledge that "neither prototypes nor dustbin matching are new in isolation" in Section 2. Of course, combining techniques can still constitute a meaningful contribution, particularly for a well-defined problem setting. However, experiments do not clearly establish why the proposed combination is necessary. For example, the authors acknowledge that "partial OT does not consistently beat forced matching". So, the authors should clarify their novelty and contribution.

W3: Repreductions and baseline comparisons

  • In Section 5, the authors state that "the published baselines are not fully aligned reproductions". Can the authors clearly explain what this means? Since FedCrossID largely combines existing techniques, a controlled comparison with recent federated ReID methods seems particularly important to establish its additional contribution.
  • However, the current results mainly demonstrate improvements over a locally implemented FedPav-style baseline, while the comparison with stronger methods remains contextual. Therefore, it is difficult to determine whether the proposed framework offers meaningful advantages over existing approaches.
Questions for authors

Q1: Would authors provide experimental results for existing federated ReID methods under the same experimental setting as FedCrossID?

Q2: Would authors further clarify the necessity of partial OT and unmatched option?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.